FLAS-Concept-46K: 46,472 Concepts for Flow-based Activation Steering

FLAS-Concept-46k is a model-agnostic corpus for training Flow-based Activation Steering (FLAS). It contains 46,472 unique concepts and approximately 2.64 million rows, each consisting of a prompt, a steered response, and a concept. The dataset was created by flas-ai and was last updated on June 2, 2026.

Use Cases

• Training concept-steering models based on the provided prompt-response-concept triples.
• Researching methods for model alignment and control using the described activation steering framework.
• Benchmarking the effectiveness of different steering approaches on a large set of concepts.

Strengths

• Contains 46,472 unique concepts for steering.
• Includes approximately 2.64 million prompt-response-concept triples.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

flas-ai

Collection Method

Likely constructed for training the FLAS method, as described in the associated paper.

Freshness

Last updated 2026-06-02 11:51:00; freshness should be verified.